• This project builds a next-word prediction app using the SwiftKey dataset (blogs and news).
• The goal is to create a lightweight, fast, and reasonably accurate prediction model using NLP tools in R.
• Due to memory/runtime limitations:
  • Only 50% of blogs and news data was used.
  • Twitter data was excluded (too noisy, informal, and interfered with clean predictions).

• The model is based on an n-gram backoff algorithm:
  • Uses trigrams → bigrams → unigrams when higher-order matches are unavailable.
  • No smoothing beyond backoff logic.
• Key Features:
  • Text is cleaned: punctuation, numbers, stopwords removed.
  • Tokenization is performed using tidytext.
  • Backoff is fast and interpretable.
• Despite its simplicity, the model achieves a perplexity of 35, indicating strong predictive performance on clean input.

• The Shiny app takes a user-entered phrase and returns predicted next word(s).
• Features:
  • Real-time predictions with a simple UI.
  • Displays the matched n-gram tier (tri-, bi-, or unigram).
• Instructions:
  1. Type in a phrase like “The sun is”.
  2. Press “Predict” — the app suggests next word(s).
  3. You can repeat and build phrases interactively.

• Runtime tradeoff: Only half the corpus used for faster loading.
• Model simplicity: No deep learning; just n-grams and backoff.
• Introduced noise: Some random n-grams added to test generalization — this slightly hindered top-1 accuracy but improved coverage.
• Exclusion of Twitter: Intentional, as its informal language hurt precision.
• However:
  • The app achieves high accuracy.
  • Lightweight, low-perplexity design suitable for mobile or embedded use.

• The app is simple, fast, and robust.
• The algorithm, though basic, delivers high predictive power.
• User experience is smooth — clean interface, responsive predictions.
• Novelty:
  • Purposefully excluded noisy data and added adversarial n-grams.
  • Achieved balance between speed and accuracy using classic methods.

I would definitely hire this person to build scalable, reliable NLP prototypes at my data science startup.